cpu_adaptive_avg_pool2d Class — pytorch Architecture

Source Code

aten/src/ATen/native/cpu/AdaptiveAvgPoolKernel.cpp lines 17–68

template <typename scalar_t, typename accscalar_t>
void cpu_adaptive_avg_pool2d(
    Tensor& output_,
    const Tensor& input_,
    IntArrayRef output_size) {
  auto input = input_.contiguous();
  auto output = output_.contiguous();

  auto input_data = input.const_data_ptr<scalar_t>();
  auto output_data = output.data_ptr<scalar_t>();

  int64_t ndim = input.ndimension();
  // treat batch size and channels as one dimension
  int64_t channels = ndim == 3 ? input.size(0) : input.size(0) * input.size(1);
  int64_t input_height = input.size(-2);
  int64_t input_width = input.size(-1);
  int64_t output_height = output_size[0];
  int64_t output_width = output_size[1];

  // parallel on dim of N, C
  at::parallel_for(0, channels, 0, [&](int64_t begin, int64_t end) {
    for (const auto c : c10::irange(begin, end)) {
      const scalar_t* input_ptr = input_data + c * input_height * input_width;
      scalar_t* output_ptr = output_data + c * output_height * output_width;

      for (const auto oh : c10::irange(output_height)) {
        int64_t ih0 = start_index(oh, output_height, input_height);
        int64_t ih1 = end_index(oh, output_height, input_height);
        int64_t kh = ih1 - ih0;

        for (const auto ow : c10::irange(output_width)) {
          int64_t iw0 = start_index(ow, output_width, input_width);
          int64_t iw1 = end_index(ow, output_width, input_width);
          int64_t kw = iw1 - iw0;

          // compute local average
          accscalar_t sum = 0;
          for (const auto ih : c10::irange(ih0, ih1)) {
            for (const auto iw : c10::irange(iw0, iw1)) {
              sum += accscalar_t(input_ptr[ih * input_width + iw]);
            }
          }
          output_ptr[oh * output_width + ow] = scalar_t(sum / kh / kw);
        }
      }
    }
  });

  if (!output_.is_contiguous()) {
    output_.copy_(output);
  }
}